hallmark

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.78). Yes: in hallmark study URL mode, the workflow performs runtime WebFetch of user-supplied URLs and parses returned HTML/CSS, producing a diagnosis report and then emitting extracted design facts into the agent’s LLM context for subsequent steps.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill's study verb explicitly WebFetches user-supplied HTTP(S) pages and uses the fetched HTML/CSS to extract "DNA" that directly controls the agent's diagnosis and subsequent build (example: the stamp example records source-url: https://example.com/), and the assets guidance also instructs runtime fetching of placeholders such as https://picsum.photos/seed/// and https://source.unsplash.com/x/?, so external URLs are fetched at runtime and can influence the agent's outputs.